115 ADAPTIVE FINITE VOLUME DISCRETIZATION OF DENSITY DRIVEN FLOWS IN POROUS MEDIA

advertisement
115
Acta Math. Univ. Comenianae
Vol. LXVII, 1(1998), pp. 115–136
ADAPTIVE FINITE VOLUME DISCRETIZATION OF
DENSITY DRIVEN FLOWS IN POROUS MEDIA
P. KNABNER, C. TAPP and K. THIELE
1. Introduction
The present paper deals with the mathematical treatment of the system of
partial differential equations, arising from the modeling of density driven flow in
porous media. A detailed description of the model is given by e.g. Leijnse [13].
The system is of a highly nonlinear and parabolic-elliptic type.
For the discretization of the problem, we use Finite Volume and/or Mixed Finite
Element Methods. For typical field applications in conjunction with safety studies
for underground waste repositories, simulations are required for large domains Ω
and a long time interval (0, T ). This requirement leads to relatively large mesh
and time step sizes because of computational limitation. This makes it necessary to consider discretization methods, where the solution fulfils proper physical
properties. For the chosen methods local conservation laws are fulfiled. Moreover
we use an error indicator to create an adaptive grid aiming at an optimal use of
computer power. By discretising the Darcy velocity artificial numerical velocities
can appear (cf. [14]). To prevent this we use a consistent approximation of the
velocity which is based on the ideas of [12].
It is well known in literature (cf. [10]) that for discontinuous and anisotropic
permeability the finite volume discretization causes some problems, because only
concentration and pressure are the primary variables and the velocity field has
to be approximated by a post-processing step. A mixed element discretization,
however, approximates the velocity directly and may gain advantages in such a
situation.
To stabilize the numerical solution w.r.t. non-physical oscillations, we use some
upwind techniques. The most common method is the “full upwind”. We also
use other upwind techniques, where the algorithm exhibits less artificial diffusion.
These techniques are described in the companion paper [11].
It should be mentioned that the theoretical analysis of the method is far from
complete. At the moment there are a lot of assumptions, conjectures, etc., which
Received November 15, 1997.
1980 Mathematics Subject Classification (1991 Revision). Primary 65M60, 65M15.
116
P. KNABNER, C. TAPP and K. THIELE
can either not be proved rigorously, or not be proved in general that the practical
experiences would indicate they hold. Nevertheless the numerical experiments
show, that with the finite volume discretization, the upwind and the adaptive grid
control based on the error indicators, we have a powerful tool for solving density
driven flow problems.
The paper is structured in the following way. In the second Section, we introduce the mathematical model of density driven flow in porous media. The
discretization in space is explained in Section 3 and 4. First we consider the
discretization of the transport equation in Section 3. For the flow equation, we
compare different types of discretizations in Section 4. An a posteriori error indicator for the error in the finite volume discretization is introduced in Section 5.
In the last Section we give some explanations of the adaptive algorithm and show
numerical results for the introduced types of discretizations and for the adaptive
algorithm.
We consider a domain Ω ∈ Rn , n = 1, 2 for any open subset ω ⊂ Ω, Lq (ω) and
l
Wq (ω) with 1 ≤ q ≤ ∞, l ∈ N denote the standard real Lebesgue- and Sobolev
spaces endowed with the norms k · k0,p,ω and k · kl,p,ω , respectively. By (·, ·) the
inner product in L2 (Ω) or L2 (Ω) × L2 (Ω) is denoted. If ω = Ω the subscript ω
is omitted. Moreover, we set H l (Ω) := W2l (Ω) and H0l (Ω) is the space of all the
functions in H l (Ω) vanishing on the boundary Γ of Ω in the usual sense of traces.
The solution space is V0 := V0 × V0 with V0 ⊂ Wql (ω), for some q > 1.
2. Mathematical Model
Given a bounded domain Ω ⊂ Rd , d = 2 or d = 3, with polyhedral Lipschitzian
boundary and an interval (0, T ), T > 0, the following nonlinear system of partial
differential equations is considered:
(1)
(2)
Φ
∂ρc
+ ∇ · (qρc − ρD∇c) = Q ,
∂t
∂ρ
+ ∇ · (ρq) = Q̃ ,
Φ
∂t
where the so-called Darcy velocity q is given by
(3)
q = −µ−1 K(∇p − ρg) .
These equations arise from the modelling of the transport of dissolved salt in flowing groundwater. The unknown functions are the mass fraction (concentration)
c = c(x, t) and the pressure p = p(x, t). A detailed discussion of the model can
be found in [13]. The equation (1) is called transport equation and equation (2)
flow equation. In particular we got the permeability tensor K : Ω → Rd×d , the
porosity φ: Ω → (0, 1), the source and sink terms Q, Q̃ : (0, T ) × Ω → R and the
ADAPTIVE FINITE VOLUME DISCRETIZATION
117
gravity g ∈ Rd . In general, the viscosity and the density µ, ρ : R → R are nonlinear
functions of c, i.e. µ = µ(c), ρ = ρ(c), with µ, ρ > 0, and the dispersion-diffusion
tensor D : Rd → Rd×d is nonlinearly depending on q.
The system is completed by the initial condition c(0) = c0 and by boundary
conditions for the unknowns pc , where the latter are of different types at different
parts of the boundary of Ω.
To the authors’ knowledge, up to now no theory of existence, uniqueness, longtime behaviour etc. of solution(s) of (1)–(3) is present. Therefore we have to make
a corresponding assumption:
There exist numbers T > 0 and q > 1 such that problem (1)–(3) possesses
c(t)
a unique solution p(t)
∈ V0 = V0 × V0 ⊂ Wq1 (Ω) × Wq1 (Ω), t ∈ (0, T ).
I.e., this is the characterizing property of the space V0 .
3. Finite Volume Discretization of the Transport Equation
We restrict the explanation of the discretization to the 2D case. The 3D case
follows analogously only by replacing triangle by tetrahedron, edge by face etc.
Starting from an admissible partition T of Ω consisting of triangles or quadrilaterals T e, e = 1, . . . , M , we define finite-dimensional spaces Vh ⊂ V0 such that the
restriction of their elements to a triangle or a quadrilaterals is a linear or bilinear
function, respectively. By N we denote the number of all vertices of the partition,
whereas I is the number of interior vertices.
Γ3im
Ωi
xm
xi
x1ij
x2ij
Γ2im
xj
Figure 1. Vertex-centered finite volume.
Following the so called vertex-centered finite volume methods, we construct a
dual mesh of control volumes Ωi , i = 1, . . . , N , by means of Donald diagrams.
Within an element T e , the barycenter is connected with the midpoints of the
element edges between the nodes xi and xj by a straight-line segment Γeij , the
midpoint of which is denoted by xeij . Here we have e ∈ Λij , where Λij denotes the
set of indices e of all elements which contain the nodes xi and xj . So we get, for
any vertex xi , a closed polygon containing it. This polygon forms the boundary
118
P. KNABNER, C. TAPP and K. THIELE
of Ωi . The numerical procedure calculates the unknowns at the vertex xi , which
is the center of the finite volume Ωi .
To derive the discretization, the equation (1) is integrated over the finite volumes Ωi , i = 1, . . . , N , and then Green’s formula is applied. The equations become
Z
φ
Ωi
∂ρc
dx +
∂t
Z
Z
Z
ρcn · q ds −
∂Ωi
ρn · D ∇c ds =
∂Ωi
Q dx.
Ωi
The arising integrals over Ωi and along the boundary ∂Ωi are approximated. In
particular, for the first one we use the quadrature rule
Z
(4)
Ωi
u dx ≈ u(xi )|Ωi |,
where |Ωi | is the area of the subdomain Ωi .
The boundary integrals we approximate by
Z
Z
ρcn·q ds−
∂Ωi
ρn·D·∇c ds ≈
∂Ωi
X
j,e
|Γeij |neij ·
1 e e
e e
e
ρ q (ci + cj ) − Kij
ρij Dij
∇ceij
2 ij ij
where |Γeij | is the length of the boundary segment Γeij . The outer normal w.r.t.
Ωi on Γeij is called neij . For the indices we got j ∈ Λi where j is the index of all
neighbour nodes xj to xi ; and e ∈ Λij . To stabilize the numerical solution w.r.t.
non-physical oscillations, we use some upwind techniques which are described by
Frolkovič [11]. With K we denote a scalar function, dependent on the local Peclet
number, with K(0) = 1 and K ≥ 1. K describes the different types of upwind.
For K ≡ 1 we have no-upwind, that means the integral is approximated by a
standard quadrature rule. In [11] also the use of 12 (ci + cj ) instead of ceij in the
approximation of the convective term will be explained.
For the sake of a short notation, we write ci := ch (xi ), pi := ph (xi ), etc.
and ceij := ch (xeij ), peij := ph (xeij ) etc. Composite functions are abbreviated by
ρh := ρ(ch ) and ρh (x) := ρ(ch (x)), ρi = ρ(ci ) and ρeij = ρ(ceij ) etc.
So we arrive at the following discrete equation:
|Ωi |φi ∂t (ρi ci ) +
X
|Γeij |neij ·
j,e
1 e e
e e
e
ρij qij (ci + cj ) − Kij
ρij Dij
∇ceij = |Ωi |Qi
2
where q is defined as in (3).
4. Discretization of the Flow Equation
We use different kinds of discretizations for the flow equation.
119
ADAPTIVE FINITE VOLUME DISCRETIZATION
4.1 Finite Volume Discretization
The flow equation can be treated in the same way as the transport equation in
the previous section. So we get for the semi-discrete problem:
Find (ch (t), ph (t)) ∈ Vh := Vh × Vh , such that for all t ∈ (0, T ) holds:
X
|Γeij |neij · ρeij qeij = |Ωi |Q̃i
(5) |Ωi |φi ∂t ρi +
j,e
(6)
|Ωi |φi ∂t (ρi ci ) +
X
j,e
|Γeij |neij
·
1 e e
e e
e
ρ q (ci + cj ) − Kij
ρij Dij
∇ceij
2 ij ij
= |Ωi |Qi
with suitable initial and boundary condition corresponding to the continuous problem. The Darcy velocity is defined in discrete form as in (3).
We make the following assumption about the existence and uniqueness of a
solution:
There exist T > 0 such that problem (5)–(6)possesses a unique solution
ch (t)
∈ Vh , t ∈ (0, T ).
ph (t)
The arising system of ordinary differential equations is solved e.g. by the implicit Euler method or a higher order time discretization.
For reason of simplicity we further restrict ourselves to the case Q, Q̃ ≡ 0,
Dirichlet boundary conditions and triangular elements.
4.2 Mixed Finite Element Discretization
In the previous section a discretization of the flow equation with the unknowns
(c, p) is formulated. In the mixed finite element method, the Darcy’s law and flow
equation are approximated individually and we get additionally the Darcy velocity
q as an unknown function.
Introducing the Hilbert space
n
o
2
H(div ; Ω) = χ ∈ L2 (Ω) | ∇ · χ ∈ L2 (Ω)
the mixed formulation of (2)–(3) is given as:
Find (q, p) ∈ H(div ; Ω) × L2 (Ω) such that
a(q, χ) − b(p, χ) = (ρg, χ) − hpD , χ · ni
∂ρ
b(ϕ, ρq) = −(φ , ϕ) + (Q̃, ϕ)
∂t
∀χ ∈ H(div ; Ω)
∀ϕ ∈ L2 (Ω)
where the bilinear forms are defined as
H(div ; Ω) × H(div ; Ω) −→ R
R
a:
(q1 , q2 )
7−→ Ω K −1 µq1 · q2 dx
2
L (Ω) × H(div ; Ω) −→ R
R
b:
(v, q)
7−→ Ω ∇ · qv dx
120
P. KNABNER, C. TAPP and K. THIELE
(·, ·) and h·, ·i denote the usual inner products in (L(Ω)2 )2 , respectively L2 (Ω) and
L2 (∂Ω).
At the moment we consider Dirichlet boundary condition, where pD is the
prescribed pressure p on the boundary.
Using the triangulation of the previous section, the following notation is introduced: The edges of an element T ∈ T are referred to as ei , 1 ≤ i ≤ 3. We denote
by E the set of edges of T and by
E 0 = E ∩ Ω,
E ∂ = E ∩ ∂Ω
the subsets of interior and boundary edges, respectively.
For D ⊆ Ω we denote the space of polynomials of degree ≤ k by Pk (D).
An approximation Vh of H(div , Ω) is given by the Raviart-Thomas space
RT0 (Ω, T ):
RT0 (Ω, T ) := {χ ∈ H(div , Ω)|χ|T ∈ RT0 (T ), T ∈ T } ,
where RT0 (T ) stands for the lowest order Raviart-Thomas element
RT0 (T ) := (P0 (T ))2 + x · P0 (T ).
Every χ ∈ RT0 (T ) is uniquely determined by the normal fluxes n · χ on the
edges ei , 1 ≤ i ≤ 3, where n denotes the outer normal w.r.t. the element T .
As a finite dimensional subspace Mh (T ) ⊂ L2 (Ω) we take the space of piecewise
constant functions:
Mh (T ) := ϕ ∈ L2 (Ω) | ϕ |T ∈ P0 (T ), T ∈ T .
Simple calculation shows
div Vh = Mh (T ).
The standard mixed discretization reads as follows:
Find (qh , ph ) ∈ Vh × Mh (T ) such that
(7)
a(qh , χh ) − b(ph , χh ) = (ρh g, χh ) − hpD , χh · ni
∂ρh
b(ϕh , ρh qh ) = − φ
, ϕh + (Q̃, ϕh )
∂t
∀χh ∈ Vh
∀ϕh ∈ Mh (T ).
This leads to an indefinite saddle point problem after linearization. To overcome
this difficulty, we use the technique of hybridisation. For more information see
[3], [9].
ADAPTIVE FINITE VOLUME DISCRETIZATION
121
Relaxing the continuity condition of the ansatz functions in RT0 (Ω, T ) we use
the non-conforming space
Vˆh = R0−1 (Ω, T ) := χ ∈ (L2 (Ω))2 | χ|T ∈ RT0 (T ), T ∈ T .
The continuity of the normal flux over element boundaries is guaranteed by Lagrange multipliers from Mh0 (E), where
Mh (E) := µh ∈ L2 (E) | µh |e ∈ P0 (e), e ∈ E
and
Mh0 (E) := {µh ∈ Mh (E) | µh |e = 0, e ⊂ ∂Ω} .
Furthermore define
MhD (E) := {µh ∈ Mh (E) | µh |e = pD , e ⊂ ∂Ω} .
Then the mixed hybrid discretization reads as:
Find (q∗h , p∗h , λh ) ∈ Vˆh × Mh (T ) × MhD (E) such that
â(q∗h ,χh ) − b̂(p∗h , χh ) + c(λh , χh ) = (ρh g, χh ) − hpD , χh · ni
∂ρh
(8)
, ϕh ) + (Q̃, ϕh )
b̂(ϕh , ρh q∗h ) = −(φ
∂t
∗
c(µh , qh ) = 0
∀χh ∈ V̂h
∀ϕh ∈ Mh (T )
∀µh ∈ Mh0 (E)
where
(9)
â :
(10)
b̂ :
(11)
c:
Vˆh × Vˆh −→ R
R
P
(q1 , q2 ) 7−→ T ∈T T K −1 µq1 · q2 dx
Mh (T ) × Vˆh −→ R
R
P
b̂(v, q)
7−→ T ∈T T ∇ · qv dx
Mh (E) × Vˆh −→ R
R
P
c(v, q)
7−→ T ∈T ∂T vq · n ds
For a solution (q∗h , p∗h , λh ) of (8), the last equation in (8) implies q∗h ∈ RT0 . If
(qh , ph ) ∈ Vh × Mh (T ) solves (7), the last equation in (8) is automatically fulfilled.
That means we drop the * in (8) and have
q∗h = qh ,
p∗h = ph .
122
P. KNABNER, C. TAPP and K. THIELE
Static condensation
Now we eliminate the unknown flux in (8). For every triangle T ∈ T there
exist 3 test functions χi ∈ Vˆh with
supp (χi ) ⊆ T.
Consider the first equation in (8) on element level. One gets (ignoring boundary
terms)
       
b1
c1
d1
q1
A  q2  =  b2  −  c2  +  d2  ,
q3
b3
c3
d3
where A is a 3 × 3 matrix with entries of the form
(A)ij = â(χi , χj ).
We define bi , ci , di as follows:
bi = b̂(ph , χi ),
ci = ĉ(λh , χi ),
di = (ρh g, χi ).
Now, inverting the matrix A, it is possible to eliminate the unknown flux in the
two last equation of (8) and solve a reduced system.
Connection to the finite volume method
In the last section we simplified our problem by eliminating the unknown flux.
Using a special quadrature rule to integrate the bilinear form a(q, χ) in (7) we end
up in a problem with one unknown per element, so we get a system of equations
like for a cell-centered finite volume method.
To do so we consider K −1 µ as a scalar and constant over the whole domain.
The triangulation fulfils a strict Delaunay property (the sum of angles opposite of
an edge ei is strictly smaller than π).
Baranger [5] proposed the following quadrature rule on elements:
Z
(12)
T
3
ph qh dx =
1X
ci φei (ph )φei (qh )
2 i=1
with
ci = cot (θi ) and θi denotes the angle opposite to the edge ei , φei (ph ) =
R
p
· n ds is the flux of ph through edge ei . (12) is exact for ph and qh piecewise
h
ei
constant functions.
ADAPTIVE FINITE VOLUME DISCRETIZATION
123
For qh ∈ RT0 , the value qh ·n is constant along the edge ei . So, for ph , qh ∈ RT0
the integral
Z
T
ph qh dx
is approximated by
3
1X
ci |ei |2 pi qi ,
2 i=1
where pi , qi are the degrees of freedom of ph , qh and |ei | is the length of the edge ei .
We use this idea to compute the integral over T in (7) for the test function
χi ∈ Vˆh with
Z
ej
χi (x) · n ds = δij .
So supp (χi ) consists of two triangles T and T 0 , see Figure 2.
1 −1
K µ(ci + c0i )φei (qh )φei (χh )
2
1
= K −1 µ(ci + c0i )qi |ei |2 .
2
a(qh , χh ) =
θi0
C0
C
θi
T0
ei
T
C (resp. C 0 is the center of the
circumscribed circle to T (resp. T 0 ))
Figure 2.
For piecewise constant p, b(p, χ) is integrated exactly. For such test function
we get
b(ph , χh ) = pT − pT 0 .
The integral (ρh g, χh ) is approximated by the same quadrature rule (12). It follows
that
1
(ρh g, χh ) = (ci + c0i )ρei g · nei |ei |2
2
where ρei is an approximation of ρ on the edge ei .
124
P. KNABNER, C. TAPP and K. THIELE
By the sine theorem, it follows for Di , the distance between C and C 0 :
ci + c0i
|ei |ni
2
1
Di = |ci + cii ||ei |
2
CC 0 =
Because of the strict Delaunay property for T, T 0 , it follows:
ci + c0i > 0 .
An extension to non-Delaunay triangulations is given in ([5]).
So, a discrete form of Darcy’s law reads as:
(13)
φei (qh ) = qi |ei |
K pT 0 − pT
=−
|ei | − ρei g · nei |ei | .
µ
Di
With the quadrature rule (4) and with Green’s formula the second equation in (7)
is given by:
(14)
|Ti |φi ∂t ρhi +
3
X
ρej φej (qh ) = |Ti |Q̃.
j=1
Now with (13) it is possible to eliminate the flux φej (qh ) in (14) and we get the
same discretization as for the well known cell-centered finite volume method.
5. Estimation of the Spatial Error
In order to derive a posteriori error estimates, we consider variational formulations of the continuous and the discrete system. In this section we consider the case
K ≡ 1, i.e. without any upwind term, but generalizations to upwind situations are
possible.
Multiplying (1) and (2) by functions r, s ∈ W0 ⊂ Wql0 (Ω), respectively, where
0
q is conjugate to q from the solution space, and setting u := (c, p) and V :=
(r, s), with V ∈ W0 := W0 × W0 , we get a variational formulation, assuming no
flow boundary conditions for simplicity both for the fluid and the mass flow or
homogeneous Dirichlet conditions by using the following forms:
ρ
a(u, V) := (ρD∇c − qρc, ∇r) +
K∇p, ∇s ,
µ
b(u, V) := (Φρc, r) + (Φρ, s),
ρ2
hf (u), Vi := (Q, r) + (Q̃, s) +
Kg, ∇s .
µ
ADAPTIVE FINITE VOLUME DISCRETIZATION
125
The variational formulation of the problem (1)–(2) reads as follows:
Find u(t) ∈ V0 , t ∈ (0, T ), such that it holds
(15)
∂
b(u, V) + a(u, V) = hf (u), Vi
∂t
c(0) = c0 .
∀V ∈ W0 ,
We get the discrete variational
form of (6)–(5) by multiplying the equations by
Vh (xi ), with Vh := srhh , Vh ∈ Wh := Wh × Wh and summing over all finite
volumes. Here Wh ⊂ W0 and consists of continuous functions.
We set:
ah (uh , Vh ) :=
I
X
rhi
i=1
−
I
X
i=1
X
neij · (qeij
j,e
shi
X
neij ·
ρeij
e
(ci + cj ) − ρeij Dij
∇ceij )|Γeij |
2
ρe
j,e
ij
K e ∇pe
µeij ij ij
|Γeij |,
bh (uh , Vh ) := (Φρh ch , rh )l + (Φρh , sh )l ,
hfh (uh ), Vh i := (Q, rh )l + (Q̃, sh )l −
I
X
shi
i=1
X
neij ·
j,e
(ρe )2
ij
e
K
g
|Γeij |,
ij
µeij
where the discrete inner product is defined by
(16)
(wh , zh )l :=
I
X
whi zhi |Ωi |.
i=1
e
Here qij
is the discrete form of (3) as discussed in Section 3.
We remember that for the indices we have j ∈ Λi where Λi is defined to be the
set of indices of all neighbouring nodes xj to xi ; and e ∈ Λij , where Λij denotes
the set of indices e of all elements which contain the nodes xi and xj . So we finally
arrive at the following discrete problem:
Find uh (t) ∈ Vh , t ∈ (0, T ), such that it holds:
(17)
∂
bh (uh , Vh ) + ah (uh , Vh ) = hfh (uh ), Vh i
∂t
ch (0) = c0h .
∀Vh ∈ Wh ,
The appearing nonlinear variational form is not well investigated. For the sake
of simplicity we consider a problem like (17), which fulfill the following assumption:
126
P. KNABNER, C. TAPP and K. THIELE
A1. There exists m0 > 0 such that V, u ∈ V0
m0 ku − VkV0 ≤ sup
W∈W0
a(u, W) − a(V, W)
.
kWkW0
A more realistic assumption is formulated in [2]. It needs some additional
properties of the asymptotic behaviour of the numerical solution, but the demands
on the bivariate form are weaker. Besides we need some additional properties
about the approximation behaviour of the partition and the chosen subspaces Vh
and some properties of the grid.
For the development of indicators for grid adaption we follow the ideas of Bieterman and Babuška [7], [8] and Angermann [2]. We separate the elliptic error by
considering an auxiliary problem. Angermann generalized the basic techniques of
a posteriori error estimation devised by Babuška and Rheinboldt [4] for nonlinear
equations. The arising local auxiliary problems are estimated further in this work
for getting easily computable quantities.
To define the auxiliary stationary problem, we introduce the notation
hf̃ , Vi := hf (uh ), Vi −
∂
b(uh , V)
∂t
and get as a new problem
(18)
a(ũ, V) = hf̃ , Vi
∀V ∈ W0 .
For the error estimation we use
(19)
ku − uh kV0 ≤ ku − ũkV0 + kũ − uh kV0 .
Using the assumption A1 we get
kũ − uh kV0 ≤
1
a(ũ, V) − a(uh , V)
sup
.
m0 V∈W0
kVkW0
With arbitrary Vh ∈ Vh we get
a(ũ, V) − a(uh , V) = a(ũ, V − Vh ) − a(uh , V − Vh ) + a(ũ, Vh ) − a(uh , Vh ).
Omitting the details, we finally obtain for the first term the relation
1
a(ũ, V − Vh ) − a(uh , V − Vh )
sup
≤ C0 η0
m0 V∈W0
kVkW0
ADAPTIVE FINITE VOLUME DISCRETIZATION
with η0q =
N
X
q
η0i
and η0i =
i=1
(0)
η̃0i :=
(1)
n X
n X
Z
hq/2
τ
τ
Z
hq/2
τ
τ
τ :τ 6⊂∂Ω
τ ∩Ωi 6=∅
(2)
η̃0i := hi
n
X
o1/q
[nτ · (qh ρh ch − ρh Dh ∇ch )]qτ dx
,
o1/q
[ρh nτ · qh ]qτ dx
,
kQ − Φ
o1/q
∂ρh ch
− ∇ · (qh ρh ch − ρh Dh ∇ch )kqLq (T )
,
∂t
kQ̃ − Φ
o1/q
∂ρh
− ∇ · (ρh qh )kqLq (T )
.
∂t
T :T ∩Ωi 6=∅
(3)
η̃0i := hi
(l)
Cl η̃0i , where Cl > 0 are certain constants and
l=0
τ :τ 6⊂∂Ω
τ ∩Ωi 6=∅
η̃0i :=
3
X
127
n
X
T :T ∩Ωi 6=∅
Here τ denotes an
n edge of oan element T from T with length hτ ,
hi := maxT :T ∩Ωi 6=∅ maxτ ⊂∂T hτ is a characteristic local mesh parameter and
[ · ]τ is the absolute value of the jump across the edge τ (with normal direction nτ ,
where nτ is the outer normal vector at τ w.r.t. T ).
To treat the second term, we start from the following decomposition:
N Z h
X
∂ρhi chi i
∂ρh ch
a(ũ, Vh ) − a(uh , Vh ) =
Q − Qi − Φ
+ Φi
rhi dx
∂t
∂t
i=1 Ωi
+
N Z h
X
Ωi
i=1
−
N
X
i=1
−
X
rhi
Q̃ − Q̃i − Φ
hZ
Ωi
∂ρhi i
∂ρh
shi dx
+ Φi
∂t
∂t
∇ · (qh ρh ch − ρh Dh ∇ch ) dx
i
e
neij · (qhΓeij ρeij ceij − ρeij Dij
∇ceij )|Γeij |
j,e
−
N
X
shi
hZ
Ωi
i=1
+
N Z
X
i=1
+
h
Ωi
X
i
ρeij neij · qeij |Γeij |
j,e
Q−Φ
i
∂ρh ch
− ∇ · (qh ρh ch − ρh Dh ∇ch ) (rh − rhi ) dx
∂t
Q̃ − Φ
i
∂ρh
− ∇ · (ρh qh ) (sh − shi ) dx
∂t
Ωi
N Z h
X
i=1
∇ · (ρh qh ) dx −
128
P. KNABNER, C. TAPP and K. THIELE
+
+
X X Z
T ∈T τ ⊂∂T
τ
T ∈T τ ⊂∂T
τ
X X Z
nτ · (qh ρh ch − ρh Dh ∇ch )rh dx
ρh nτ · qh sh dx .
The first four terms are estimated straightforward via Hölder’s inequality and
nP
o1/q0
I
q0
by making use of the equivalence of the norm
|w
|
|Ω
|
(cf. (16))
hi
i
i=1
0
on the components of Vh with the Lq (Ω)-norm. The estimates of the fifth and
the sixth term are in fact estimates of the lumping error. The last two terms are
estimated by means of scaled trace inequalities.
For the other term in (19) we assume, that ∃δ ∈ (0, T ) ∃CT > 0 ∃κ > 0
∀t ∈ (δ, T ):
ku − ũkV0 ≤ CT hκ η0
Collecting the components of the estimators and putting similar terms together,
we finally obtain the estimate
ku(t) − uh (t)kV0 ≤
7
1 X
Cl η̃l
m0
with η̃lq :=
l=0
and
η̃0i
η̃1i
η̃2i
i=1
Z ∂ρhi chi
∂ρh ch
dx ,
:= |Ωi |
+ Φi
Q − Qi − Φ
∂t
∂t
Ω
Z i 0 ∂ρh
∂ρhi
−1/q := |Ωi |
Q̃ − Q̃i − Φ
+ Φi
dx ,
∂t
∂t
Ωi
Z
0
:= |Ωi |−1/q ∇ · (qh ρh ch − ρh Dh ∇ch ) dx
X
Ωi
neij · (qeij
j,e
η̃4i
η̃liq
−1/q0
−
η̃3i
N
X
ρeij
e
(ci + cj ) − ρeij Dij
∇ceij )|Γeij |,
2
Z
X
−1/q e e
e
e := |Ωi |
∇ · (ρh qh ) dx −
ρij nij · qij |Γij | ,
Ωi
j,e
Z
n X
o1/q
:=
hq/2
[nτ · (qh ρh ch − ρh Dh ∇ch )]qτ dx
,
τ
0
τ
τ :τ ∈∂Ω
/
τ ∩Ωi 6=∅
η̃5i :=
n X
hq/2
τ
τ :τ ∈∂Ω
/
τ ∩Ωi 6=∅
η̃6i := hi
n
X
T :T ∩Ωi 6=∅
Z
τ
o1/q
[ρh nτ ·qh ]qτ dx
,
kQ − Φ
o1/q
∂ρh ch
− ∇ · (qh ρh ch − ρh Dh ∇ch )kqLq (T )
,
∂t
ADAPTIVE FINITE VOLUME DISCRETIZATION
η̃7i := hi
n
X
T :T ∩Ωi 6=∅
kQ̃ − Φ
129
o1/q
∂ρh
.
− ∇ · (ρh qh )kqLq (T )
∂t
In general, the constants C0 , . . . , C7 depend on global parameters associated
with the bivariate forms a, ah , b, bh , (·, ·) and with the family of partitions of Ω (cf.
also [1] for the situation in the case of a linear convection-diffusion problem). For
the numerical experiments we take q = 2.
6. Adaptive Algorithm and Numerical Experiments
Two kinds of numerical experiments were done. First we consider the adaptive
algorithm for the finite volume discretization of flow and transport equation. The
main attention is payed to the behaviour of the adaptive procedure. In the second part of experiments we compare the finite volume discretization of the flow
equation with the mixed finite element discretization.
We restrict ourselves to 2-d problems. The adaptive procedure for mesh control
utilizes the error indicators which have been developed in the previous section 5.
Here the basic strategy is to get an equidistribution of the following local indicators:
7
X
η̃li
, i = 1, . . . , N .
ηi :=
η̃l
l=0
In this way we guarantee that all indicators are dimension-free and have an equal
influence on the new grid. This representation of indicators is supported by experiments which show that the results are not very sensitive w.r.t. a moderate
variation of the weighting factors.
In order to be able to start from a rather coarse initial grid, the first step of the
algorithm is performed by means of a re-starting procedure, i.e. after an evaluation
of the indicators corresponding to the actual spatial mesh, it is decided whether
the second time step has to be carried out or the mesh has to be refined and then
the first time step is to be repeated.
All further time steps are passed through only once. During the computations
the adaptive algorithm is used only eventually, i.e. the grid is kept fixed for a
number of time steps.
In the computations the software package UG from the University of Stuttgart
is used. The arising nonlinear system is linearized via Newton’s method and the
linear equations are solved by a multigrid method. The size of the time step is
chosen by the nonlinear solver of UG. For further information about UG and the
built-in solvers see [6].
A widely accepted test example is so-called Elder problem. There Ω is a rectangular domain (0, 600) ×(0, 150) in the x-z-plane. The data are: K ≡ 4.845 ·10−13 I
(where I is the identity matrix), D ≡ 3.565 · 10−6 I, µ ≡ 10−3 , φ ≡ 1/10,
130
P. KNABNER, C. TAPP and K. THIELE
0
Q ≡ Q̃ ≡ 0, g = −9.81
. The density is of the form ρ = 1000 + 200 c. Furthermore, c0 = 0. The boundary conditions are as follows: c ≡ 1 for z = 150,
x ∈ [150, 450] and c ≡ 0 elsewhere, n · q = 0 on the whole boundary. Since the
pressure is determined
up to an additive constant, we additionally require p = 0
0
.
in the point 150
As a second example we consider a modification of the Elder problem. Here
within the rectangular domain (0, 600) × (0, 300) we have several subdomains. We
have jumps of permeability over the boundaries of the subdomains, particular:
K ≡ 4.845 · 10−13 I in the subdomain Ω1 . K ≡ 4.845 · 10−16 I elsewhere, for the
adaptive procedure, and
K=
4.845 · 10−13
0
0
4.845 · 10−16
for the mixed elements. All other coefficients and the boundary conditions are like
in the Elder problem.
6.1 Numericals Results for the Adaptive Algorithm
Because no exact solution is known for the Elder problem, we look for the
plausibility of the adaptive procedure. In the Figures 3 and 4 the concentration
contour lines and the adaptive grids for the time 1.5 and 3.5 years are shown. The
grids consist of 6 234 and 6 910 nodes, respectively. We see that the evolution of
Figure 3. Concentration contours and adaptive grid at t = 1.5 years.
ADAPTIVE FINITE VOLUME DISCRETIZATION
131
Figure 4. Concentration contours and adaptive grid at t = 3.5 years.
the grid closely follows the evolution of the concentration profile. The algorithm
resolves well the most important areas of the computational problem. The solution is compared with a results on a uniform grid of 16 384 nodes in Figure 5. The
qualitative behaviour of the solution is the same with significantly less computational costs for the adaptive procedure. The computational costs of the indicators
are comparatively low because no additional boundary value problems are to be
solved. For the specific problem, only 10% of the total computing time is spent to
compute the estimator and for grid-refinement.
Figure 5. Concentration contours for a uniform grid at t = 3.8 years.
For the modified Elder problem we look for the behaviour of the adaptive algorithm near inner boundaries. It is well known in literature [10] that the jump
of the coefficients at the inner boundaries may cause problems. In Figure 6 the
132
P. KNABNER, C. TAPP and K. THIELE
Figure 6. Velocity field, adaptive grid and concentration contours at t = 10 years.
ADAPTIVE FINITE VOLUME DISCRETIZATION
133
contour lines of the concentration, the velocity field and the adaptive grid are presented for the time 10 years. The grid consists of 16 157 nodes. Like in the Elder
problem we notice that the grid follows the contour lines of the concentration.
Moreover, at the inner boundaries of the domain the finer grid sizes appear. So
the area, where we suppose the most problems in calculation, are of finer grid size.
6.2 Numericals Results for the Mixed Finite Element Method
The purpose of this section is to discuss and to compare the differences between
the finite volume method (FV) and the mixed finite element method (MFE).
In FV one obtains a linear approximation of the pressure p. The calculation of
the Darcy flux q with the consistent velocity approximation results in a cellwise
constant q. In MFE the pressure p and the velocity is approximated individually.
This leads to a piecewise constant pressure in each element (or a non-conform linear
pressure in each element [3]) and a piecewise linear approximation for velocity in
the elements.
This difference becomes clear by the Figures 7 and 8. They show the velocity
field for the modified Elder example after the first time step.
Figure 7. Velocity field from FV.
134
P. KNABNER, C. TAPP and K. THIELE
Figure 8. Velocity field from MFE.
Figure 9. Velocity field and concentration from FV.
ADAPTIVE FINITE VOLUME DISCRETIZATION
135
Figure 10. Velocity field and concentration from MFE.
A further difference between FV and MFE is the mass conservation property.
The FV fulfils this property on the finite volumes Ωi and the MFE on elements.
It is easy to see that for the problems with strong discontinuous or anisotropic
permeability the results of the MFE are more realistic as for FV on the same grid.
References
1. Angermann L., An a posteriori estimation for the solution of elliptic boundary value problems by means of upwind fem, IMA Journal of Numerical Analysis 12 (1992), 201–215.
2.
, A posteriori error estimates for approximate solutions of nonlinear equations with
weakly stable operators, Preprint 209, Institut für Angewandte Mathematik, Universität
Erlangen-Nürnberg, 1996.
3. Arnold D. N. and Brezzi F., Mixed and nonconforming finite element methods: Implementation, postprocessing and error estimates, M 2 AN 19(1) (1985), 7–32.
4. Babuška I. and Rheinboldt W. C., Error estimates for adaptive finite element computations,
SIAM Journal on Numerical Analysis 15(4) (1978), 736–754.
5. Baranger J., Maitre J. and Oudin F., Connection between finite volume and mixed finite
element method, M 2 AN 30(4) (1996), 445–465.
136
P. KNABNER, C. TAPP and K. THIELE
6. Bastian P. and Wittum G., Adaptive multigrid methods: The UG concept, Adaptive methods-Algorithms, theory and applications, Proceeding of the Ninth GAMM Seminar, Kiel,
January 22.-24. 1993 (W. Hackbusch and G. Wittum, eds.), Vieweg,, Braunschweig-Wiesbaden, 1994, pp. 17–37.
7. Bieterman M. and Babuška I., The finite element methode for parabolic equations i. a posteriori error estimation, Numerische Mathematik 40 (1982), 339–371.
8.
, The finite element methode for parabolic equations i. a posteriori error estimation
and adaptive approach, Numerische Mathematik 40 (1982), 373–406.
9. Brezzi F. and Fortin M., Mixed and hybrid finite element methods, Springer-Verlag, 1991.
10. Durlowsky L., Accuracy of mixed and control volume finite element approximation to darcy
velocity and related quantities, Water Resources Research 30(4) (1994), 965–973.
11. Frolkovič P., Maximum principle and local mass balance for numerical solutions of transport
equations coupled with variable density flow, Acta Mathematica Universitatis Comenianae
LXVII(1) (1998), 137–157.
12. Knabner P. and Frolkovič P., Consistent velocity approximations in finite element or volume
discretizations of density driven flow in porous media, Computational Methods in Water
Resources XI, Cancun. (A. A. Aldama et. al., ed.), Computational Mechanics Publications.
Southampton, 1996, pp. 83–100.
13. Leijnse A., Three-dimensional modeling of coupled flow and transport in porous media, PhD
thesis, University of Notre Dame, Indiana, 1992.
14. Voss C. I. and Souza W. R., Variable density flow and solute transport simulation of regional
aquifers containing a narrow freshwater-saltwater transition zone, Water Resources Research
23 (1987), 1851–1866.
P. Knabner, Universität Erlangen-Nürnberg, Institut für Angewandte Mathematik I, Martensstraße 3, D–91058 Erlangen, Germany, e-mail: knabner@am.uni-erlangen.de
C. Tapp, Universität Erlangen-Nürnberg, Institut für Angewandte Mathematik I, Martensstraße 3, D–91058 Erlangen, Germany, e-mail: tapp@am.uni-erlangen.de
K. Thiele, Universität Erlangen-Nürnberg, Institut für Angewandte Mathematik I, Martensstraße 3, D–91058 Erlangen, Germany, e-mail: thiele@am.uni-erlangen.de
Download